In order to improve the accuracy of fault diagnosis in analog circuits, aimed at the instability structure and error transferring of the existing directed acyclic graph support vector machine (DAGSVM), a novel approach based on fault distinguish degree to construct DAGSVM is proposed. According to the fault information acquired from all of the testable points, this paper defines the concept of fault distinguish degree, and takes this as a basis to optimize the topology of the DAGSVM to eliminate the inherent instability of DAGSVM structure. For this reason, there is a stable and quite good accuracy of diagnosis. The experimental results show that this method improves obviously diagnosis accuracy compared with "1 vs 1",SVM, and traditional DAGSVM, and simultaneously the method can be used for reference in analog circuit fault diagnosis.